On Wed, Jul 08, 2020 at 11:12:02PM -0700, Kees Cook wrote:

On Tue, Jul 07, 2020 at 03:30:49PM +0200, Christian Brauner wrote:

quoted

Hm, maybe change that description to sm like:

[...]

Cool, yeah. Thanks! I've tweaked it a little more

quoted

quoted

+	/* 24 is original sizeof(struct seccomp_notif_addfd) */
+	if (size < 24 || size >= PAGE_SIZE)
+		return -EINVAL;

Hm, so maybe add the following:

#define SECCOMP_NOTIFY_ADDFD_VER0 24
#define SECCOMP_NOTIFY_ADDFD_LATEST SECCOMP_NOTIFY_ADDFD_VER0

and then place:

BUILD_BUG_ON(sizeof(struct seccomp_notify_addfd) < SECCOMP_NOTIFY_ADDFD_VER0);
BUILD_BUG_ON(sizeof(struct open_how) != SECCOMP_NOTIFY_ADDFD_LATEST);

Yes, good idea (BTW, did the EA syscall docs land?)

I'll be giving a kernel summit talk about extensible syscalls to come to
some agreement on a few things. After this we'll update the doc patch
we have now and merge it. :)

I've made these SECCOMP_NOTIFY_ADDFD_SIZE_* to match your examples below
(i.e.  I added "SIZE" to what you suggested above).

Yup, sounds good!

quoted

somewhere which is what we do for clone3(), openat2() and others to
catch build-time nonsense.

include/uapi/linux/perf_event.h:#define PERF_ATTR_SIZE_VER0     64      /* sizeof first published struct */
include/uapi/linux/sched.h:#define CLONE_ARGS_SIZE_VER0 64 /* sizeof first published struct */
include/uapi/linux/sched/types.h:#define SCHED_ATTR_SIZE_VER0   48      /* sizeof first published struct */
include/linux/fcntl.h:#define OPEN_HOW_SIZE_VER0        24 /* sizeof first published struct */
include/linux/fcntl.h:#define OPEN_HOW_SIZE_LATEST      OPEN_HOW_SIZE_VER0

The ..._SIZE_VER0 and ...LATEST stuff doesn't seem useful to export via
UAPI. Above, 2 of the 3 export to uapi. Is there a specific rationale
for which should and which shouldn't?

I think openat2() just didn't think it was useful. I find them helpful
because I often update codebase to the newest struct I know about:

struct clone_args {
	__aligned_u64 flags;
	__aligned_u64 pidfd;
	__aligned_u64 child_tid;
	__aligned_u64 parent_tid;
	__aligned_u64 exit_signal;
	__aligned_u64 stack;
	__aligned_u64 stack_size;
	__aligned_u64 tls;
/* CLONE_ARGS_SIZE_VER0 64 */
	__aligned_u64 set_tid;
	__aligned_u64 set_tid_size;
/* CLONE_ARGS_SIZE_VER1 80 */
	__aligned_u64 cgroup;
/* CLONE_ARGS_SIZE_VER2 88 */
};

But bumping it means I can't use:

clone3(&clone_args, sizeof(clone));

everywhere in the codebase because I'm fscking over everyone on older
kernels now. :)

Soin various parts of the codebase I will just use:

clone3(&clone_args, CLONE_ARGS_SIZE_VER0);

because I don't care about any of the additional features and I don't
need the kernel to copy any of the other stuff. Then in other parts of
the codebase I want to set_tid so I use:

clone3(&clone_args, CLONE_ARGS_SIZE_VER1);

This way I can also set "templates", i.e.

struct clone_args clone_template1 = {
	.flags		|= CLONE_CLEAR_SIGHAND,
	.exit_signal	= SIGCHLD,
	.set_tid	= 1000,
	.set_tid_size	= 1,
};

and then use the same struct for:

clone3(&clone_template1, CLONE_ARGS_SIZE_VER0);
clone3(&clone_template1, CLONE_ARGS_SIZE_VER1);

Whereas sizeof(clone_template1) would always give me
CLONE_ARGS_SIZE_VER2.

Christian